This article shows how a high-quality filtering architecture for targetted documents can be achieved by adapting the intex system to the domain (i.e. ‘economic intelligence’), when expressed in finite state automata filters. A push system named corail (‘request composition by linguistic intelligent agents’) is described, developed by a consortium comprising Thomson-CSF Corporate Laboratory (senior partner), Informatique-CDC, the University of Paris X/CRIS, and the ladl. Both practical and theoretical issues regarding the use of finite state automata are discussed, as for example the notion of the ‘automaton classes’ which are to be used as a paraphrasing agent in the process of document filtering. This project can be considered as an example of how linguistics and computer science can interact in anatural language processing task to meet high-quality standards, namely avoiding noise and reducing silence, thus offering an alternative solution to traditional push systems based on simple keyword-matching. A prototype of the corail system is now operational at Thomson-CSF Central Laboratory and Informatique-CDC.